Omnilingual ASR Corpus

The Omnilingual ASR Corpus is a speech dataset published by Meta's FAIR team as part of its wider Omnilingual ASR release. It gathers transcribed spontaneous speech for roughly 348 languages that are poorly served by existing speech technology, spanning writing systems including Latin, Arabic and Devanagari. In total it holds around 3,350 hours of recordings, each paired with a human transcription, collected in partnership with researchers and community organisations across Africa, Asia and beyond.

Access is straightforward: the corpus is hosted on Hugging Face as a public dataset stored in Parquet (and optimised-Parquet) files, so you can stream it or pull it in bulk with the datasets library. Because audio and transcription travel together, it is well suited to training and evaluating automatic speech recognition, and to fine-tuning multilingual audio models. For a retrieval-oriented or text AI workflow, the transcripts themselves are the useful part: a rare pool of natural, conversational text in languages that are otherwise thin on the ground, which can seed multilingual knowledge bases or augment low-resource training sets.

Watch-outs are mainly about fit and scale. Coverage is deliberately skewed towards low-resource languages, so per-language volume varies and this is not a general-purpose English or high-resource corpus. The models and code in the same release carry an Apache 2.0 licence, but the corpus is CC BY 4.0, so attribution to Meta is required and you should keep the two licences separate in your own records. As spontaneous speech, transcripts include disfluencies and informal register, which is realistic but may need cleaning for some uses. If you need broader multilingual speech, Mozilla Common Voice and the wider Common Voice ecosystem are natural companions already worth cataloguing alongside this.